AI Agents: How Do We Take Back Control? | Full Documentary

Everyone everywhere is talking about AI agents.

Systems that run continuously, replace entire layers of human work, and make decisions on our behalf.

Their potential is impossible to ignore, and they're quickly moving from theory into practice.

I'm Alex [music] Kantrowitz, and I spent my career interviewing the people building this technology, people like Sam Altman and Dario Amodei.

And I think there's something different about this shift.

AI agents might look like just the latest tool, but they don't behave like the tools we're used to.

Are we giving up too much control?

What's at risk if we do? And are we prepared for a world where humans aren't the only ones making the decisions?

To find out, I'm speaking to the people at the forefront of researching, deploying, and trying to secure these systems. We have not seen a technology transformation [music]

at least in my lifetime, like this.

It's moving at an absolutely blistering pace. [music]

pace. [music] The more tools we give them, the more autonomy we give them, the more powerful they get.

When you're doing it on a scale of one, two, three agents, it might sound [music] great. When you start getting to

[music] great. When you start getting to 10s, 15s, your productivity is going to go off a cliff.

People are actually sabotaging their own AI agents.

Oh, absolutely.

[music] To understand where this is heading, I wanted to start with what AI agents are, how they work, and what they could become.

[music] Ramesh Raskar is a researcher at MIT Media Lab, where he's [music] been exploring how AI agents interact with one another, and what that might mean as they become

more autonomous. [music]

more autonomous. [music] Ramesh, good to see you.

Wonderful to see you.

So, AI agents is this technology that everybody is speaking about, but it could use a definition. What is an AI agent?

An AI agent fundamentally has four parts. It has some intelligence

parts. It has some intelligence and that can be, you know, a language model that is accessing and knows about the world.

Second is that it has access some tools.

It can, you know, search the web, can do calculation, do simulations.

Uh it has accumulated memory of what it has done in the past. Uh and most importantly, it has a Rolodex. You know,

it has connection to other agents.

How is this different from previous technologies?

The key distinction is the autonomy and action. You know, agents can work while

action. You know, agents can work while we sleep. And agents can connect with

we sleep. And agents can connect with other agents. And the more tools you

other agents. And the more tools you give them, the more autonomy you give them, you know, the more powerful they get.

So, for AI agents to work, you have to give them a lot of trust, right? You

have to sometimes give them access to your entire computer. Do you see a situation where people will say, "I trust these agents enough"?

That's the dilemma because autonomy comes because of intelligence and intelligence comes because of autonomy.

Of how do we allow access to very private and sensitive data that allows us to create our own personal agent that's very, very smart

that will take actions and negotiate on our behalf. So, there's a lot of trust. It's not just about the data, but

trust. It's not just about the data, but also trust in their actions.

If you would look at the internet between like 0% agents and to 100% of their capabilities, where would you say we are right now?

I would say we're kind of in the, you know, 0.001 phase of what agents can do.

I want to hear a little bit about this looks like at the 100% Right.

from the optimistic side. What could

things look like if everything goes well?

So, in the first phase we are saying let's improve productivity through agents, right? Agents for X. The agent

agents, right? Agents for X. The agent

is almost like an assistant, like a co-pilot. But when you go to the second

co-pilot. But when you go to the second phase, the transformation, you know, X for agents, services for agents. That's

where we start transforming health care because we're building infrastructure for agents to talk to patients and run clinical trials and invent drugs and so on. But then, you

know, the amazing phase is going to be the third one, the disruption where you know, agents are creating their own teams. They're creating their own organization. They're creating their own

organization. They're creating their own stock markets, you know, their own justice system. And I think that phase

justice system. And I think that phase if we can make sure that humans are still in control I think we can.

Everything's going to change.

What could happen if this goes wrong?

The unfortunate part is that there's a very good chance this will go wrong because it's moving so fast. And um most of the decisions are being made by a handful of companies

or handful of countries. The fear is that, you know, there's exactly one AI agent for finance. There's exactly

one AI agent for teachers. You know,

there's exactly one AI agent for mental health. And then everyone just uses

health. And then everyone just uses that. And uh you know, all of us just

that. And uh you know, all of us just become kind of these intelligent slaves where even if you don't lose our jobs it's complete homogeneity, complete centralization.

So, we are we go to work, we pay for the agent and then the agent is basically making our decisions for us.

Yeah, go to sleep. And we continue to feed the machine. So, everybody's just trying to survive.

That's bleak.

Yeah, it's just bleak. It's it's quite likely. Unfortunately.

likely. Unfortunately.

this?

Uh right now we have we have no mechanisms. So, a lot of people are talking about AI safety but nobody, Alex, nobody's talking about AI agent safety yet. You know, that's my

mission right now.

If you do it now, it's much easier.

Before AI agents are highly diffused in the society.

And if we do that, the potential is immense.

You know, to cure diseases, to have education, have better democracy.

So, you're convinced that the technology is there. The real question is how it's

is there. The real question is how it's implemented.

I think the guardrails to observe, monitor, audit this humans of agents is going to be very critical. So, if we do it right, and if we have to do it now,

there's a chance that we will have a you know, a very prosperous next 50 years.

If If we can do it right.

Can I have a black coffee, please?

Thank you.

Okay.

Thank you. Have a good one. Thank you so much.

I'm thinking a little bit differently about AI [music] agents now. You know,

the common perception is that an AI agent is just this one agent that does work for you or a company. But Ramesh

made something really clear. It's not

about one agent. It's going to be groups of agents that are going to work for us together, communicate with each other.

And I'm really looking forward to figuring out how this technology can be implemented safely, effectively, and in a way that everybody can benefit from, and whether that's even possible.

From what I've seen, this is being driven as much by competition as by opportunity.

If you're not moving quickly, you're already falling behind.

Which is why the [music] most enthusiastic adopters so far are in the private sector.

Sharon Guy has seen companies [music] grapple with deploying AI agents up close. So, I wanted to know what she

close. So, I wanted to know what she thinks about the potential and the reality of AI agents in business.

Sharon, why would a company roll out an AI agent at all? What is the value that they would get from doing it right?

Agents at the end of the day are are 24/7. They work tirelessly. They will

24/7. They work tirelessly. They will

never complain. There is no company politics and it is a more objective rational system that can take over work for you.

Right now, it feels like AI agent is still more of a buzzword than a current reality in many enterprise settings.

Um is there a single thing or a group of things holding this technology back?

I think first there's um an emotional thing. So, I think, you know, in the US,

thing. So, I think, you know, in the US, our jobs are part of our identity and when you threaten that, it's not going to cause a smooth roll out. We've seen a lot of uh folks actually sabotage their

AI projects.

Wait, people are actually sabotaging the roll out of AI agents?

Oh, absolutely. Why would they be incentivized to encourage a smooth roll out? I think a lot of companies have

out? I think a lot of companies have sort of chased the shiny new tech and say, oh, I think this might be useful for our employees. Whereas, I think the right way to look at roll out is

probably from a person's role itself.

Isn't there this thing happening within enterprises where they're reluctant to roll it out because you have to give immense amounts of data and access to these agents?

Yeah, how can we trust it um when the technology is so young and there's already so many cases from um Replit and Cursor and uh Claude itself where it's

gone rogue and deleted an entire company's database. So, um when a a

company's database. So, um when a a larger enterprise sees that happening to a startup, they worry. There's a lot of risk involved and there's a lot of

um things that you have to clean up if it doesn't go right.

you balance the power that you can get from AI agents then with the risks that they might go rogue and mess your stuff up?

Uh at the end of the day, it's human loop. It's having an overseer look at

loop. It's having an overseer look at this output and seeing if that is uh usable.

Don't you think it's funny that we're talking about humans directing AI agents where maybe one day they'll direct us.

That's entirely possible. Just imagine

if I created an AI agent uh I've taught this agent everything I know. I leave

and you're the new hire. This AI agent knows so much more than you do. It's

more quote-unquote senior than you. And

so maybe for a short while you will report into it.

I've had some bad bosses in the past, so maybe I wouldn't mind trying out working for an AI agent.

Yeah, maybe we all will.

It's clear that this is a technology being pushed hard by AI native startups.

But for AI agents to truly reshape [music] work, they'll have to be adopted by the world's biggest companies.

So, how close are we to that tipping point?

To find out, [music] I caught up with Ambika Rajagopal, Michelin's Group Chief Data and AI Officer.

[music] The perfect person to explain how this is playing out at enterprise scale.

[music] Hi Ambika, great to see you.

Thank you for having me.

So, Michelin is a company that is working on tires, new composites, and some restaurant ratings. Where within

Michelin are you seeing real uses of AI agents?

We have about 100 use cases that we're looking at. Um I'll give you three. So,

looking at. Um I'll give you three. So,

number one is uh you know, within our supply chain where we have AI models uh deployed in multiple geographies uh to do our demand forecast. Uh number two,

we have AI in our plants that are using computer vision to identify tire defects. And number three, of course, is

defects. And number three, of course, is to listen to our customers. We have AI models that listen to weak signals that come to us from multiple sources on emerging needs, feedback from our

customers, which we then react to.

How important is AI's ability to access tools and use different programs on its own to the operation that you're putting into place at Michelin?

Everything in terms of AI being able to model uncertainty, etc., is predicated upon, you know, having the data, information, semantic models, the data

layer. All of this really accessible in

layer. All of this really accessible in a way that is intelligible by agents.

Our systems are really not built for agents to be first-class citizens, first-class consumers of any of this information. At the end of the day, you

information. At the end of the day, you know, is the data layer really ready to really enable these agents to deliver the value that they can. And this I think is the real challenge that

enterprises are facing, including us.

So, how do you decide how far to put your trust in these agents and what to let them access?

If I create an agent today, you know, the agent can then have only access to the data, the tools, the applications that I as a user have access to. But

let's say I want to roll out that agent to 100 people in my team. This aspect of handling, you know, rule-based access security for agents, we have found to be

not as mature and in response to this, we have invested in our own understanding around AI agents and AI security in general. And we have a

rigorous security review of any agent that's actually going into deployment.

who say I'm going to give this a shot, what are the risks and challenges you think they might encounter as they go forward?

The number one aspect is, you know, for the AI teams and the business teams to work closely together to understand the technology better, understand the business problem more accurately, and

put in the time with each other to build trust mutually as well as in the technology. However, the far more

technology. However, the far more difficult part is to identify processes, problems, areas of complexity within the business processes of an enterprise

where AI can really bring value.

It's clear, Ambika, that you are approaching this technology responsibly, but it's a big economy out there in a big world. Do you have some fear about

big world. Do you have some fear about how the world will approach this and whether that responsibility will extend beyond your walls?

I do believe that enterprises that have advanced understanding of this technology may be more advanced than, you know, regulatory bodies and governments at

this time have definite responsibility to ensure that it is deployed in a responsible way. For me, a lot of it

responsible way. For me, a lot of it really, you know, comes down to transparency and explainability.

Ambika, thank you so much. Great

speaking with you.

Thank you.

Ambika's work at Michelin is a glimpse of where the rest of the economy is heading. And if enterprises like hers

heading. And if enterprises like hers are already this deep into deployment, the rest of business probably [music] isn't far behind. So, as this technology spreads, the question becomes,

are we ready for the consequences?

[music] [music] I wanted to speak to someone who's seen how decisions around technology are made at the highest levels [music] of government.

Teresa Payton was the Chief Information Officer of the White House, [music] responsible for some of the most sensitive systems in the world.

I'm curious to know what she thinks about the potential of AI agents, the risks [music] they might pose, and how prepared the government really is.

So, Teresa, it's clear that there's a lot of excitement around AI agents, including from yourself.

But then there is a lot of questions about what they're going to do. Are

these things ready to be trusted with real responsibility?

I think you can give the technology real responsibility if you have the right governance and guardrails in place. I'm a daily user of AI,

place. I'm a daily user of AI, but at the same time, it's a technology that I would say, you know how you hire somebody right out of either high school or college or

graduate school, and they're bright and they have tons of promise.

So, you just want to make sure you put the right management around them so that they don't make decisions based on a lack of professional experience. That's what I would say

experience. That's what I would say about AI right now. It's an incredibly brilliant, tons of promise, new hire.

So, given your experience, how do you think the government will approach policy towards AI agents?

I think the challenge that our government has is that our legislative process is still very much reflect technology that's sort of like horse and cart.

We're passing laws for yesterday's challenges and problems. We have not seen a technology transformation, at least in my lifetime, like this. We have

the consumerization of AI algorithms, access to data lakes, access to agents that prior to now

was only in the hands of engineers. And

now we are at a stage where we're making the technology actually generate technology on its own.

When you have agents building other agents, you could potentially have an explosively bad situation, and it could be too late to clean it up.

We're already seeing in labs where AI agents, if they have to make a choice to be turned off or save a human life, that they say let the human die.

Wait, they do what?

Let the human die, because they they don't want to be turned off.

That's messed up.

We already have nation-states and cybercriminal syndicates who are using this technology to steal intellectual property and to steal money. But also, I don't want to

steal money. But also, I don't want to stop the technology from happening, because I see where AI is headed right now, it's going to be a great equalizer if we

get it right. I think that's worth holding our elected officials and holding private sector enterprises accountable to say, you have to do more to ensure human flourishing.

Industry would say, you know what, Teresa, good point. But, if we were to legislate now as this thing starts to take off, there's a good chance that it will prevent all of the positive benefits

from happening.

How do you respond?

What does winning look like at the end here?

So, we don't put governance and guardrails up, and then we win.

What do we have left? Are we going to be proud of it?

Do you think Washington is well-equipped to handle this?

No, and I I and I don't mean that out of any disrespect for the people. There

shouldn't be a massive bill of legislation. I think we have to attack

legislation. I think we have to attack it sort of one problem at a time. The

first challenge that we have is incentive alignment. So, how do you

incentive alignment. So, how do you incentivize these technology companies to think about us? Because at the end of the day, things are not going well for us.

Who are they selling their services to?

All right.

And I I think one of the challenges that we have right now is the inaction of passing new laws is an action.

And it's an action that has really disastrous and consequences. So, we have 36 months or less to actually impact the next 50 years.

Does that scare you?

If we don't move quickly on governance and guardrails or big tech coming out and saying, "We're going to do this."

It does scare me.

Well, that was really tough to hear.

You hear about the government's inability to tackle this AI moment [music] and you think, "Well, on the inside they're probably much better prepared." But, this was the former CIO

prepared." But, this was the former CIO of the White House telling us that she believes that the government is wholly unprepared for this moment. [music] And

you think about the fast-moving technology that we see in action and the government that's [music] standing still and it really makes you wonder what those next critical 36 months are going to look like and who

[music] is going to lead on this.

As AI agents become more deeply embedded in how we work and live, how do we get the benefits of these systems without losing control?

Someone who I hope can shed some light on this is Rory Blundell, CEO of Gravity, whose work focuses on securing AI agents at scale.

Hi Rory, good to see you.

Hey Alex.

What's the state of the AI agent rollout today?

In actual fact, we've done some research fairly recently that shows about 80% of CIOs and CTOs are starting to deploy, but there's a

very small proportion of those same groups of people that we've interviewed that really feel that they have control.

Only about 10% give or take.

80% of CIOs have rolled out some form of AI agent technology, but 10% think they have it under control.

Yeah, that's crazy.

It's It's completely crazy, you're right.

I can't help but feel that in some ways we're screwed no matter what. I mean, if you think about the government, the government can't handle this stuff. If

you think about the volume that we're going to have coming through with agents, it's going to be extraordinary.

Am I off base by at least having some feeling of skepticism or hopelessness when I think about this?

No, but I feel honestly much more optimistic actually.

Really?

And yes, I do. I have an 18-month-old son, and it really changed my perspective on the world and on what I'm trying to leave behind. And what I

realized is there are ways to harness this, but the nature and the power of these particular technologies, I think can be overwhelming for people to comprehend its scale.

How can a company possibly think they have a chance of securing this technology when they have these swarms of agents that have new capabilities today that they didn't have yesterday?

Well, I believe the answer is a little more simple than people feel because they see the scale of the challenge and they think there's no way that we can deal with this.

My recommendation is that all of the different architectural components of an AI ecosystem go through and are controlled by single pane of glass.

What's a single pane of glass?

If you look at an AI ecosystem for a typical enterprise organization, what's it comprised of?

So, you'll have a piece of software, let's call it the agent, that has a set of instructions. It typically

of instructions. It typically communicates with one of three things.

It will communicate with its brain, the large language model, small language model, Claude, Gemini. It will also communicate with things that we think of as tools. The third thing that it might

as tools. The third thing that it might interact with are other agents. What we

believe is required and what Gravity does is we sit on all three of those interactions.

The single pane of glass reports all of that ecosystem centrally for you. So,

you have one place that you can control, govern observe.

Rory, aren't there going to be two sets of companies? Like one set that puts in

of companies? Like one set that puts in all these governance practices into place, and then a set that just says, "Screw it. We're going full agent. We'll

"Screw it. We're going full agent. We'll

worry about the safeguards later."

Yeah, that will happen, but I don't agree that actually they will be the ones that will succeed. I will actually think in the long run those ones will fail. When you're doing it on a scale of

fail. When you're doing it on a scale of one, two, three agents, it might sound great. When you start getting to 10s,

great. When you start getting to 10s, 15s, your productivity is going to go off a cliff.

So, there is this fallacy that in order to handle the the bad sides of AI agents, there might be this magical technology bullet that comes in.

But, I think what you're arguing is it's really more of a new security mindset that has to come in first before the technology can do the work.

I think they work in parallel. It's not

necessarily just a technological thing.

This is about here. It's about taking human beings on the journey from being the human as the builder to the human as the architect.

You have a child who's 18 months old.

Mhm.

When you think about where the world is heading, what do you hope for them?

I hope that if we heed the learnings of previous generations, that he would benefit from all of those massive massive outcomes that AI can

bring for us.

That'd be pretty cool.

I hope so.

Thank you, Rory.

You're very welcome.

As we come to the end of our conversations, I'm left thinking that AI agents are happening whether we like it or not. With that being said, we have an

or not. With that being said, we have an opportunity to control the way that it is deployed and the way that [music] it influences our society and we better get invested because if we let it go the bad

way there are going to be some serious consequences.

But if we do it right, it can be an important [music] and beneficial technology for our society and really that begins today.

[music] [music] [music]